Most AI agents fail in production because their execution loops lack governance rather than intelligence. This article presents a production-ready architecture covering step budgets, state management, tool routing, evidence verification, auditability, testing, and escalation patterns that make AI agents reliable, controllable, and suitable for enterprise and regulated environments.